Machine-learning assisted assessment of CO₂ leakage through adjacent wells in geological carbon storage

Geological CO₂ storage in deep saline aquifers represents a central component of long-term climate mitigation strategies. Despite their large storage potential, concerns remain regarding CO₂ leakage, particularly through adjacent wells, which constitute one of the most critical pathways compromising storage integrity and long-term storage effectiveness. Although numerical reservoir simulators are capable of representing complex multiphase flow behavior, their high computational cost limits large-scale uncertainty analysis. This limitation motivates the need for computationally efficient yet physically interpretable approaches for leakage risk assessment.

This study develops an integrated workflow that combines large-scale numerical simulation and machine learning to jointly evaluate CO₂ storage capacity and leakage risk in saline aquifers with existing adjacent wells. A total of approximately 7,000 simulations are computed using CMG-GEM to represent geological and operational conditions, including variations in reservoir and aquifer properties (e.g., permeability and porosity), caprock permeability, distance between the injection well and the adjacent well (either abandoned or monitoring well), and adjacent-well damage severity.

Artificial neural network models are trained to predict total securely stored CO₂ and cumulative leaked CO₂ mass, showing near-perfect agreement with numerical simulation results (R² ≈ 0.99). In parallel, a random forest model is implemented to classify leakage behavior into low, high, and extreme risk regimes based on leakage fraction thresholds commonly adopted in CCS studies. Lastly, model interpretability is assessed using Morris screening and partial dependence plots to identify the dominant controls commanding storage and leakage behavior. The results indicate that reservoir porosity is the primary control on secure CO₂ storage capacity, whereas leakage behavior is mainly influenced by the distance between the injection well and the adjacent well, followed by reservoir permeability and well damage severity. On the other hand, caprock and aquifer properties exhibit a comparatively minor influence. 

The proposed framework enables rapid screening of a large number of potential storage site configurations that would otherwise be computationally impractical to evaluate using conventional numerical simulations. By providing reliable estimates of storage capacity and leakage risk at low computational cost, the framework supports practical, physics-informed decision-making during the early stages of CCS project planning, particularly for site selection and injection strategy design.

 

Acknowledgement: This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2025-25436989, RS-2025-24803244).